Transcript Bringing the User Back into Scheduling: Two Case Studies of

S
Bringing the User Back into
Scheduling:
Two Case Studies of Interaction with Intelligent
Scheduling Assistants
www.sri.com
Hypothesis: Real-World domains are better served by collaborative approaches
that involve the user in the scheduling process than full-automation
PTIME: Personalized Time Management
0 Elicit Preferences
3 Choose Request
1 Elicit Request
Mixed Initiative Collaboration
via
1. Intuitive elicitation
of requests
4 Learn Preference Model
0
2. Unobtrusive adaptation to
User’s personal
preferences
Elicit
Preferenc
e
1
Elicit
Request
2
User Study
1.One-on-one vs.
group
2.Mandatory vs.
optional
3.Fixed vs. floating
1.Importance
2.Urgency
3.Interest (topic)
4.Relationships
5.Perturbation
6.Stability
7.Preferences
Y3 Focus in Bold
Perception
of Event
Process
FOCI
1.Walk-in
2.Constraint
satisfaction
3.Iterative refinement
4.Add-in
Need
Factors
Preferences
3
Unobtrusive refinement of the preference model through
natural interaction with the user
Choose
Schedule
Constraint Reasoner
Objective Function
An SVM (QP-based) learner generalizes these specific
comparisons to produce a second linear function
4
SVM
Learner
The two functions are combined in PLIANT to form the
preference model for the user
This dynamic model is used to derive the objective function by
which scheduling options are evaluated, presented to the user, and
scheduled
Preference Model
Scheduling
Process
Elicitation
Process
1.Coordinating
groups of busy
people
2.Intelligent
reminders
3.Transparency
4.Control
Schedule
Options/Selection
Constraints
LP-based Learning
SVM Learner
aizi + aijzizj  bixi
PLIANT
1.Meeting specific, e.g., day time
2.Calendar-wide e.g., density
3.How to relax meeting constraints
4.How to relax calendar-wide constraints
Preference Model
bixi + (1-) cixi
Pisces: Large-Scale Logistics Scheduling
•
Problem
Spec.
(Goals)
User
•
•
User-enabled exploration of
solution space
User manipulation of solution
and solution criteria
Multiple and diagnostic views
Pisces UI
Problem
Session
Management
Schedule
Session
State
Scheduling
Engine
• Continuous, incremental schedule improvement
– Rapid response, anytime solution generation
• Support for complex evaluation functions
– Resource balance, stability, activities included
• Support for specialized query directives
– Reduce peak in certain regions; Resource Flattening
cixi
Mixed Initiative
Collaboration via
1. Direct manipulation of
the scheduling problem
2. Facilitated exploration
of the solution space
– Keep activity in schedule; Fix an activity in time
– Shift Left and Right
Pauline Berry, Bart Peintner and Neil Yorke-Smith
Artificial Intelligence Center, SRI International, 333 Ravenswood Ave, Menlo Park, CA 94025
{berry, peintner, nysmith}@ai.sri.com
Objective Function